On Dependency Analysis via Contractions and Weighted FSTs

Arc contractions in syntactic dependency graphs can be used to decide which graphs are trees. The paper observes that these contractions can be expressed with weighted finite-state transducers (weighted FST) that operate on stringencoded trees. The observation gives rise to a finite-state parsing algorithm that computes the parse forest and extracts the best parses from it. The algorithm is customizable to functional and bilexical dependency parsing, and it can be extended to non-projective parsing via a multi-planar encoding with prior results on high recall. Our experiments support an analysis of projective parsing according to which the worst-case time complexity of the algorithm is quadratic to the sentence length, and linear to the overlapping arcs and the number of functional categories of the arcs. The results suggest several interesting directions towards efficient and highprecision dependency parsing that takes advantage of the flexibility and the demonstrated ambiguity-packing capacity of such a parser.